GPT-Rosalind: OpenAI's Frontier Reasoning Model for Life Sciences

GPT-Rosalind: OpenAI's Frontier Reasoning Model for Life Sciences

TL;DR

OpenAI has launched GPT-Rosalind, a frontier reasoning model specifically optimized for the life sciences, including biology, drug discovery, and translational medicine. The model is designed to accelerate the early stages of scientific discovery by improving evidence synthesis, hypothesis generation, and experimental planning.

Optimized for Scientific Workflows

GPT-Rosalind is engineered to handle the complexity of modern scientific research, which often involves fragmented workflows across large volumes of literature, specialized databases, and experimental data. The model series is optimized for reasoning over molecules, proteins, genes, pathways, and disease-relevant biology.

Key capabilities include:

  • Multi-step workflow support: Enhanced effectiveness in literature reviews, sequence-to-function interpretation, experimental planning, and data analysis.
  • Tool and database integration: Improved ability to select and use computational tools and domain-specific databases to augment reasoning.
  • Biochemical reasoning: Specialized training to handle chemical reaction mechanisms, protein structure, mutation effects, and phylogenetic interpretation of DNA sequences.

Performance Benchmarks

GPT-Rosalind was evaluated against several public and private benchmarks to measure its core reasoning and its ability to support real-world research tasks.

Public Benchmarks

  • BixBench: GPT-Rosalind achieved leading performance among models with published scores on this bioinformatics and data analysis benchmark.
  • LABBench2: The model outperformed GPT-5.4 on 6 out of 11 tasks, with the most significant improvement noted in "CloningQA," a task requiring the end-to-end design of DNA and enzyme reagents for molecular cloning protocols.

Expert-Level Performance

In partnership with Dyno Therapeutics, OpenAI evaluated the model on RNA sequence-to-function prediction and generation using unpublished sequences. When used within the Codex app, the best-of-ten model submissions performed as follows:

  • Prediction task: Ranked above the 95th percentile of human experts in the AI-bio field.
  • Sequence generation task: Ranked around the 84th percentile of human experts.

Ecosystem and Tooling

To facilitate the integration of the model into research, OpenAI has released a Life Sciences research plugin for Codex, available on GitHub.

This plugin serves as an orchestration layer providing access to over 50 public multi-omics databases, literature sources, and biology tools. It supports repeatable workflows such as sequence search, protein structure lookup, and public dataset discovery. While the Life Sciences model is reserved for qualified users, the plugin is available for all users to use with mainline OpenAI models.

Deployment and Trusted Access

Due to the potential for biological misuse, GPT-Rosalind is being deployed via a trusted-access program for qualified Enterprise customers in the United States. Access is granted based on three principles:

  1. Beneficial Use: Organizations must conduct legitimate scientific research with a clear public benefit.
  2. Governance: Organizations must maintain strong safety oversight and misuse-prevention controls.
  3. Security: Access must be restricted to approved users within secure, enterprise-grade environments.

OpenAI is currently collaborating with organizations such as Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific to apply the model to discovery workflows.

Future Directions

This release marks the first in a series of life sciences models. OpenAI plans to continue expanding the model's biochemical reasoning and support for long-horizon, tool-heavy workflows. This includes ongoing partnerships with national laboratories, such as Los Alamos National Laboratory, to explore AI-guided protein and catalyst design and the modification of biological structures to improve functional properties.

Sources